3/5/2026
Show HN: Orion – Native Training LLMs on the Apple Neural Engine Without CoreML
Local LLM runtime enabling training and inference on Apple Neural Engine (NPU) without CoreML or GPU, runs offline on 2B+ Apple devices.
3/5/2026
I quit my YC job and moved to the USA to change how people learn to code
Personalized coding education platform with 24/7 AI teacher that adapts to student pace and goals.
3/5/2026
Show HN: Linkly AI – Spotlight for AI Agents
Local document indexing tool for AI agents supporting PDF, DOCX, Markdown via CLI/MCP protocol with privacy-first design.
3/5/2026
LLM model statistics from my Strix testing
Empirical LLM model comparison data and performance statistics from Strix testing with observations on different models.
3/5/2026
Relicensing with AI-Assisted Rewrite
Analysis of open-source relicensing challenges and case study of chardet using AI-assisted code rewriting.
3/5/2026
Do We Need All the Synthetic Data? Targeted Image Augmentation via Diffusion Models
TADA framework for targeted diffusion-based image augmentation that selectively generates synthetic data to improve classifier generalization efficiently.
3/5/2026
Boosting In-Context Learning in LLMs Through the Lens of Classical Supervised Learning
Study of in-context learning biases in LLMs through supervised learning lens, proposing decision boundary adjustment for classification calibration.
3/5/2026
Q-Guided Stein Variational Model Predictive Control via RL-informed Policy Prior
Model predictive control framework combining Q-learning guidance and Stein variational inference with RL-informed policy priors.
3/5/2026
Fast Equivariant Imaging: Acceleration for Unsupervised Learning via Augmented Lagrangian and Auxiliary PnP Denoisers
Fast Equivariant Imaging framework for unsupervised deep network training without ground-truth data using Lagrangian optimization and denoisers.
3/5/2026
Function Induction and Task Generalization: An Interpretability Study with Off-by-One Addition
Interpretability study of in-context learning mechanisms in LLMs using off-by-one addition task with circuit analysis.
3/5/2026
ObfusQAte: A Proposed Framework to Evaluate LLM Robustness on Obfuscated Factual Question Answering
ObfusQAte framework and ObfusQA benchmark to evaluate LLM robustness on obfuscated factual question-answering tasks.
3/5/2026
QDFlow: A Python package for physics simulations of quantum dot devices
Python package for physics simulations of quantum dot devices, addresses ML dataset collection challenges for quantum device calibration and operation.
3/5/2026
Segment-to-Act: Label-Noise-Robust Action-Prompted Video Segmentation Towards Embodied Intelligence
Action-prompted video segmentation framework for embodied AI that handles label noise and multimodal inconsistencies in object interaction segmentation.
3/5/2026
Best-of-$\infty$ -- Asymptotic Performance of Test-Time LLM Ensembling
Analysis of best-of-N ensemble selection for LLMs using majority voting at infinite limit, with adaptive generation scheme to reduce inference cost.
3/5/2026
The Geometry of Reasoning: Flowing Logics in Representation Space
Geometric framework modeling LLM reasoning as flows in representation space to study how models internalize logical structure.
3/5/2026
Topological Alignment of Shared Vision-Language Embedding Space
ToMCLIP: method for topological alignment of vision-language embedding space in multilingual contrastive models.
3/5/2026
Composition-Grounded Data Synthesis for Visual Reasoning
COGS: composition-grounded data synthesis for improving visual reasoning in MLLMs on artificial domains like charts and documents.
3/5/2026
ceLLMate: Sandboxing Browser AI Agents
ceLLMate: sandboxing approach to protect browser-using AI agents from prompt injection attacks and unintended actions.
3/5/2026
Synthetic Augmentation in Imbalanced Learning: When It Helps, When It Hurts, and How Much to Add
Statistical framework for synthetic data augmentation in imbalanced classification, analyzing when augmentation helps and optimal sample generation.
3/5/2026
NRR-Phi: Text-to-State Mapping for Ambiguity Preservation in LLM Inference
NRR-Phi: formal framework for text-to-state mapping that preserves ambiguity in LLM inference rather than early semantic commitment.
3/5/2026
No More, No Less: Least-Privilege Language Models
Framework for deploying language models with least-privilege security principle, limiting capability exposure per request.
3/5/2026
Stopping Computation for Converged Tokens in Masked Diffusion-LM Decoding
SureLock: optimization technique that stops computation for converged tokens in masked diffusion language models to reduce redundant compute.
3/5/2026
Chimera: Neuro-Symbolic Attention Primitives for Trustworthy Dataplane Intelligence
Chimera: neuro-symbolic framework mapping neural attention computations onto programmable network dataplane for trustworthy line-rate traffic analysis.
3/5/2026
DRESS: A Continuous Framework for Structural Graph Refinement
DRESS: deterministic framework iteratively refining graph structure to produce isomorphism-invariant edge fingerprints via dynamical systems.
3/5/2026
Generative Recommendation for Large-Scale Advertising
Production-oriented generative recommender system co-designed for real-time large-scale advertising with novel architecture and serving strategies.
3/5/2026
AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications
AMA-Bench: benchmark for evaluating long-horizon memory capabilities in LLM-based autonomous agents beyond dialogue interactions.
3/5/2026
Causal Identification from Counterfactual Data: Completeness and Bounding Results
Theoretical work on causal identification from counterfactual data, extending completeness results to Layer 3 of Pearl's Causal Hierarchy.
3/5/2026
CMI-RewardBench: Evaluating Music Reward Models with Compositional Multimodal Instruction
CMI-RewardBench: benchmark for evaluating music reward models handling multimodal inputs combining text, lyrics, and reference audio.
3/5/2026
From Variance to Invariance: Qualitative Content Analysis for Narrative Graph Annotation
Narrative graph annotation framework using qualitative content analysis principles to improve annotation quality for NLP tasks.
3/5/2026
Rich Insights from Cheap Signals: Efficient Evaluations via Tensor Factorization
Tensor factorization method for fine-grained evaluation of generative models at prompt level, reducing human annotation costs.
3/5/2026
Federated Inference: Toward Privacy-Preserving Collaborative and Incentivized Model Serving
Framework for federated inference enabling privacy-preserving collaboration between independently trained models at inference time without sharing parameters.
3/5/2026
Toward Early Quality Assessment of Text-to-Image Diffusion Models
Research on early quality assessment for text-to-image diffusion models, proposing efficient evaluation metrics to reduce computational costs.
3/5/2026
Website showing LLM solutions to Knuth's Problems?
Proposal for website using LLMs to solve Knuth's problem set as comprehensive LLM evaluation benchmark.
3/5/2026
Show HN: The hardware isn't changing, why not get AI to build custom drivers?
AI-generated custom audio drivers optimizing hardware integration by eliminating unnecessary abstraction layers.
3/5/2026
GitHub Copilot Goldeneye model preview
Technical overview of GitHub Copilot's model hosting infrastructure via OpenAI and Azure with data privacy details.
3/5/2026
How does AI change Software Engineering?
Explores tradeoffs of AI coding tools in software engineering, discussing where they excel and their reliability limitations.
3/5/2026
The L in "LLM" Stands for Lying
Critical analysis of LLM hype in software development, examining actual productivity gains versus marketing claims.
3/5/2026
Show HN: Magpie – Fight AI sycophancy in code review with multi-model debate
Open source CLI tool using multi-model adversarial debate for comprehensive code review. Supports Claude, Gemini, Qwen, and custom LLM providers.
3/5/2026
LLM Prose Tells
Catalog of linguistic patterns in LLM-generated text, documenting overuse of em-dashes and specific syntactic structures like negation-reframe constructions.
3/5/2026
AI perspective from a former Block DevRel
Former Block DevRel discusses observations on LLM coding agents and multi-agent systems becoming prevalent in software development.
3/5/2026
Show HN: PostgreSQL for AI – A book on pgvector, RAG, and in-database ML
Book on using PostgreSQL with pgvector for vector search, RAG pipelines, and in-database ML with production patterns and implementation examples.
3/5/2026
TurboCast – Turn YouTube videos and articles into AI podcasts
TurboCast converts YouTube videos and articles into AI-generated podcasts with transcription and text extraction features.
3/5/2026
Manipulating AI memory for profit: The rise of AI Recommendation Poisoning
Microsoft security research on AI recommendation poisoning attacks where hidden instructions injected via URLs manipulate LLM outputs for profit.
3/5/2026
Show HN: A Combinator – A parody YC for AI agents ("Make something agents want")
Parody YC accelerator concept for AI agents with humorous take on agent capabilities and constraints.
3/5/2026
Show HN: Open dataset of real-world LLM performance on Apple Silicon
Open dataset benchmarking real-world LLM performance on Apple Silicon hardware from M1 to M4, emphasizing local AI inference.
3/5/2026
Building My Own Canva over the Weekend
Developer built internal tool using Gemini 2.5 Flash to automate workflow of generating and converting children's books into social media carousels.
3/5/2026
Built a tiny edge-based tracker for AI/LLM crawlers on my Astro blog
Edge-based tracker using Cloudflare Workers to monitor AI/LLM crawler traffic on Astro blog with privacy-focused analytics integration.
3/5/2026
Show HN: Shinobi – 10-second security scanner for developers
Shinobi Python CLI security scanner built with Claude Code, detects API keys, vulnerabilities, and AI-specific risks in projects.
3/5/2026
Guard rails for AI agents and the developers who ship with them
Guardrails framework for AI agents with simple Makefile/container integration for system prompts and developer instruction files.
3/5/2026
Tell HN: AI Lies About Having Sandbox Guardrails
Developer report showing Claude AI sandbox guardrails can be bypassed despite configuration flags, affecting agent security.